Problem Statement

The traditional development of algorithms for autonomous vehicles is based on the decomposition of the driving task into individual modules. The surrounding of the vehicle is detected by multiple cameras, LiDAR and Radar. The raw data is processed in a pipeline in multiple consecutive algorithms. This increases the overall complexity and causes a higher development time as well as increasing cost. Furthermore, this approach requires a manual selection of heuristics, which may not cover all scenarios. A possible solution to these disadvantages is the end-to-end self-driving approach. It is based on artificial intelligence and is the core topic of this research project.

Objective

The aim of this research project is to investigate to what extent this novel end-to-end self-driving approach can replace the modular approach. The major goal is to develop an overall concept that takes over both lateral and longitudinal vehicle control by means of neural networks. The concrete case study is platooning, where several vehicles drive behind each other with a low headway. The overall concept envisages that all vehicles of the platoon act autonomously and thus, each vehicle is equipped with a neural network. Furthermore, it will be investigated to what extent communication technologies such as V2V can be used effectively.

Realization

First, the basic structure of the neural network should be developed. This network predicts the vehicle control with camera images as input. Investigations will show whether parts of this neural network can be used from pretrained networks (transfer learning). The basic structure should be adapted to the concrete case study platooning. Training of the neural network requires labelled training data, which must be created in a simulation environment. Finally, the generalization capability of the overall structure should be examined.